AMD’s Ryzen 9000-series CPUs, codenamed “Granite Ridge”, are manufactured by TSMC in part utilising their latest 4nm 4NX process. Setting aside architectural improvements, a die shrink should allow the silicon to operate with lower voltages and at higher frequencies, meaningfully improving power efficiency at key performance bands. AMD has chosen to re-use the cIOD, manufactured using TSMC's 6nm process, from the 7000-series rather than substantially revise its design ahead of new memory standards being released.

The Zen 5 architecture incorporates key improvements over Zen 4 while adhering to the same CCD + cIOD chiplet layout that has proved so successful since its debut in 2019’s 3000-series. Each 9000-series CPU can feature up-to 2 CCDs for compute and a single large IOD which houses functionality that benefits less from the cutting-edge process node. These include PCIe signalling, the memory controller, iGPU and other aspects of the SoC.

Ryzen 9000-series CPUs conform to the AM5 socket standard and are thus backwards compatible with AMD 600-series motherboards. AM5-compatible CPU coolers remain compatible with the new processors as there are no mechanical differences to be accounted for. TDP for each of the classes of processor has changed however, with both 9700X and 9600X slimmed down to a 65W TDP envelope from 105W. In theory, this should mean that a wider selection of CPU coolers will be able to cope with the cooling requirements of these ostensibly mid-range processors.



Internal AMD testing indicates that Zen 5 can boast a double-digit percentage IPC improvement over Zen 4 averaged across selected workloads, continuing Ryzen’s impressive track record of generational performance uplift. To achieve these IPC improvements Zen 5 incorporated a swathe of tweaks, changes and additions over Zen 4. Chiefly, AMD focussed on increasing data bandwidth and L1 per-core cache structures, but a complete list of changes touches on most aspects of the CPUs function.

Particular emphasis was placed on the AVX-512 datapath for this update, which was doubled to 512-bit, and more support for the bfloat16 floating point data structure, both of which enable greater acceleration of workloads aligned with machine learning. Conventional workloads will benefit from a more advanced Branch Predictor that performs two predictions per cycle, leveraging the larger L1 data cache and improved L2 bandwidth.

The launch 9000-series CPUs are all equipped with a nominal integrated 2 CU (128-shader) iGPU based on RDNA 2. They offer hardware audio and video encode/decode (including AV1 hardware decode) and display-out if supported by the motherboard, but these desktop processors are designed to be paired with a discrete GPU when gaming. Integrated graphics are however excellent troubleshooting features and in some cases can be leveraged to accelerate relevant workloads that might ordinally be offloaded to the dGPU.

The Granite Ridge SoC also has the following features as standard:



Additional I/O functionality in the form of ancillary PCIe lanes, peripheral expansion and storage (SATA and M.2) is offered by the motherboard chipset with configuration determined by the vendor.



Overclocking

Overclocking as a test is beyond the scope of this review but we would be remiss to not mention that the old Curve Optimiser tool which allowed Ryzen 7000-series users to undervolt the CPU in a manner that was PMFW/PBO aware has been joined by a new tool. The ‘Curve Shaper’ tool can assign voltage margins in up-to 15 different frequency-temperature bands, allowing for finer control for voltage adjustment in bands that are stable and unstable. Curve Shaper tables apply across all cores, and can also be shifted as a group by Curve Optimiser.